Skip to content
Main Site News Console

Anthropic's IPO Roadshow Video

· Simon Willison Translated
技巧LLM

Anthropic’s IPO Roadshow Video. Anthropic launched their IPO roadshow yesterday, and the roadshow video contains some previously undisclosed details about how their business operates.

Co-founder Tom Brown’s segment on their “Approach to compute” is particularly interesting, as he describes how Anthropic trains and deploys Claude across AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs.

I’m not sure why there isn’t an official transcript of the video, but here is a transcript I extracted using MacWhisper. Some highlights:

We now train and deploy Claude across all four major chip families. This is possible because we made an early, unconventional bet on a multi-chip strategy. At the time, many believed that running AI models across different chips wasn’t worth the effort. Each chip requires a tailored software and inference stack, which is expensive and difficult to build. We’ve been doing this for several years now, and we’ve gotten very good at it. […]

As an example, this year we started using AMD GPUs for the first time. In the past, bringing up a new chip platform took dozens of engineers several months. This time, we let Mythos take a crack at it. With almost no human intervention, it got Claude up and running on AMD in just three days. Claude ported or wrote nearly 100% of the kernels. […]

The second driver of our compute advantage is the rate at which we discover “compute multipliers.” We focus on total effective compute, which is the product of physical capacity and the compute multipliers we discover. Every multiplier we find makes our pool of total effective compute larger. For instance, suppose you need 100 units of compute to train a model of a given intelligence level. A 4x compute multiplier means you can train that exact same model using only 25 units of compute. Alternatively, you can use those 100 physical units of compute to train an even smarter model that previously would have required 400 units.

Compute multipliers create a competitive advantage for us that is difficult to replicate. They compound on one another. Every new multiplier we discover positively impacts the ones that came before it. They stack. We believe this process could accelerate even further as we approach recursive self-improvement. Claude will be able to discover compute multipliers entirely on its own, causing the number of compute multipliers to grow faster and faster.

We caught a glimpse of Anthropic’s multi-chip strategy back in September in their postmortem of three recent issues, where they described bugs that exclusively affected their TPU deployments.